Study On Array Speech Recognition Under Complex Environment

Speech recognition belongs to the field of artificial intelligence and speech processing, itcan make machines to understand human language, and take appropriate action in accordancewith the person’s commands. Single channel speech recognition has been developing rapidlyat present, and has good recognition effection. However, it has some disadvantages of poorflexibility, need to wear a microphone, limiting the speaker’s activities. The microphone arraycan overcome the disadvantages of single channel speech recognition, therefore, in recentyears the microphone array speech recognition has become a research focus.Based on the domestic and international research progress of speech recognitiontechnology, this paper analyses the current speech recognition problem. This paper expoundsthe basic theories of speech signal preprocessing, including sampling, quantization, framing,windowing, endpoint detection and so on, and analyses the Mel frequency cepstrumcoefficient in detail. This paper also studies the three basic algorithms of the HMM model, thechoice of primitives and the number of state determine in speech recognition, meanwhilegives the problems and solutions of HMM model in application.For the effection of single channel speech recognition in a real environment is not ideal,this paper first puts forward a kind of array speech recognition method based on channelselection. For actual closed environment, the paper constructed the correlation matrix arraysignal after delay compensation, then subspace decomposition. In the signal subspace, usingchannel selection method based on multiple cross-correlation, remove the channel that hasless correlation. Select a group channels that has maximum cross-correlation, and then getoutput signal through the beamformer. Last, the speech recognition result obtained by thespeech recognizer. On this basis, the speech recognition is not only a question of signalprocessing, but a model discriminant problem. Therefore, the information in the speechrecognition system are used in the front end of the array processing, and using conjugategradient algorithm to find the maximum likelihood probability of correct assuming, thenapplied to the speech recognizer. Simulation results show that the method not only reduces theamount of microphone and computation, but also enhances useful information in speechrecognition to improve the rate of recognition. Meanwhile it has better robustness in complexacoustic environment.